Kaminario playing 3D flash chippery doo-dah with its arrays

Storing hot and cold

El Reg: What are your views on flash tiering with, say, fast flash used for primary data and 3D TLC (slow) flash used for cold data?

Fienblit: We already implemented in our software the technology to separate between hot and cold data. We are using this technology today to effectively decrease the over-provisioning of the overall flash in our system and for further reducing the write amplification.

We are planning to utilise this IP in the future for utilising lower-endurance drives and to tier between two types of Flash drives within the same node. This will help us to improve density per node in a very cost-efficient way.

El Reg: Do you see Kaminario entering into an integration deal with a server manufacturer to arrive at an integrated server/networking/storage system (like VSPEX and FlexPod)?

Fienblit: This is something we are examining with our server partners, as it brings simplicity of implementation. In general, we believe that separating between compute and storage allows the customer to build a very cost-efficient solution.

This is a much more cost-efficient solution compared with hyper-converged solutions, that in many implementations over-provision capacity or performance based on the mixture required for the use case.

El Reg: How do you see Kaminario responding or reacting to the server-SAN, hyper-converged infrastructure appliance ideas, such as EVO:RAIL, that are spreading currently?

Fienblit: We expect to see hyper-converged solutions gain some market share in the low-end storage array market because of simplicity, but the cost of these solutions will significantly limit the adoption for many other use cases. There are two significant limitations in hyper-converged that move the cost up compared to storage arrays in a converged infrastructure:

• Inability to separate between capacity and performance means high over-provisioning on compute or capacity for a typical use case.
• Very low storage utilisation in hyper-converged solutions driven by high capacity loss related to the usage of a data mirroring approach, and full node reserve required for supporting fault tolerance.

Capacity utilisation in a typical hyper-converged system is between 25 and 40 per cent, which is a significantly lower utilisation compared with the 87.5 per cent of Kaminario K-RAID, our high-performance, dual-parity RAID.

We believe that an integrated server/networking/storage solution is a more cost-efficient solution and therefore will dominate the market.